A large proportion of what you are told about food rests on people reconstructing six months of ordinary eating from memory, over the phone, to a stranger who is evaluating them. That is not a reason for nihilism — but it is a reason to ask what kind of evidence you are looking at before you ask what it says.
Imagine a court case where the only evidence is a phone call.
Someone rings the witness and asks: over the past six months, how many biscuits did you eat in an average week? And roughly what size were they? And while we’re at it, how much coffee — cups per day, and how big were the cups?
The witness does their best. They’re not lying. But they’re reconstructing six months of ordinary life from memory, on the phone, to a stranger who is clearly evaluating them. Nobody in that situation reports the Tuesday night when they ate most of a packet standing up at the kitchen counter.
You would not convict anyone on that. You would not even open a case.
And yet an enormous proportion of what you are told about food — this reduces your risk of that, this many portions per day, this food is linked to that disease — rests on exactly that. Food frequency questionnaires. Twenty-four-hour recalls. Memory, filtered through the very human wish to appear reasonable.
I don’t say this to be nihilistic about nutrition science. I say it because the strength of the evidence behind a claim varies enormously, and almost nobody tells you which kind you’re looking at.
The obvious question is why anyone would build a field on recalled diet.
Because the alternatives are brutally expensive.
There are objective methods. Urinary nitrogen tells you protein intake without asking. Plasma carotenoids indicate fruit and vegetable consumption. Doubly labelled water measures energy expenditure directly. There’s a method where participants prepare two identical meals — eat one, send the other to a laboratory. Extremely accurate, and completely impractical at scale.
These approaches cost several times what a questionnaire-based study costs. So the questionnaire persists, not because anyone thinks it’s good, but because it’s what the budget allows. And a large study built on a weak instrument still produces a publishable number with a confidence interval around it, which looks like precision.
A very large study can be built on a very soft measurement, and the size does not fix the softness.
Here’s the sequence that produces most of what you’ll read this week.
1. A compound is tested in a dish. Cells behave interestingly. The paper is published, appropriately hedged, in a specialist journal.
2. The university press office writes a release. The hedges thin out.
3. A news outlet covers the release rather than the paper. The word “may” does a lot of load-bearing work in the third paragraph and disappears from the headline.
4. A supplement company launches a product citing the research.
5. Three years later, a human trial finds nothing. That paper gets a fraction of the coverage, because a null result isn’t a story.
The cycle works because correction doesn’t travel. Nobody shares the article saying the thing they shared last year turned out not to hold.
And the dose problem is the part that never survives the translation. Compounds that do remarkable things in a dish are frequently present in food at a fraction of the concentration used, with a fraction of the absorption. The mouse got an amount per kilogram that a person could not consume from food in any realistic quantity — and mice aren’t people anyway.
Not scams. Not useless. Just promoted well beyond what human trials support.
Green tea extract at high dose. The population data on drinking green tea is one thing. Concentrated EGCG capsules are another. Large trials haven’t delivered the cancer prevention the marketing implies, metabolic effects are modest at best, and high-dose extracts have a documented association with liver injury serious enough that supplement safety bodies have flagged it.
Curcumin supplements. The animal data is genuinely spectacular. The human data is inconsistent, the trials are small and often unblinded, and curcumin is famously poorly absorbed — which is why formulations add other compounds to force absorption, at which point you’re no longer testing turmeric. Concentrated extracts also carry liver injury reports.
Superfood berry blends. No trials showing disease prevention. Some short-term movement in antioxidant markers, which is not the same as a health outcome. The marketing-to-evidence ratio here is among the worst in the category.
Coconut oil as a heart-healthy fat. It raises LDL cholesterol to a similar degree as butter. The medium-chain triglyceride argument doesn’t hold at the amounts people actually eat. This one is instructive because the mechanism sounded plausible and the trials simply didn’t agree.
Juice cleanses. No human evidence for detoxification beyond what the liver and kidneys already do continuously. Weight lost is calorie restriction, and returns. Low protein and fibre means some of what’s lost is muscle.
Because the point isn’t that nothing works.
Vitamin D, in people who are deficient. The VITAL trial — nearly 26,000 participants over five years — found no reduction in overall cancer incidence or major cardiovascular events. But the benefit signal sits with those who were deficient at baseline, which is precisely the group a supplement should help. That’s not a failure of vitamin D. It’s a demonstration that supplementing a replete person achieves little.
Magnesium. Consistent, if modest, reductions in systolic blood pressure across multiple trials. It also participates in vitamin D metabolism, which is why the two are worth considering together rather than separately.
Omega-3, in people who don’t eat fish. This is the clearest illustration in the whole field. VITAL found no overall cardiovascular benefit at a gram a day — but a prespecified subgroup with low habitual fish intake did benefit. Same supplement, same trial, opposite conclusions depending on who you were before you started.
Soluble fibre. Reliable effects on LDL and glycaemic control. More dependable than most supplements, partly because the outcome is directly measurable rather than inferred.
Specific probiotic strains for specific conditions. The evidence is strain-specific and condition-specific, and it does not generalise to “probiotics are good for you.” Worth knowing: many commercial products don’t contain the strains used in the research, at the counts used in the research.
Coffee is the perfect case, because it exposes the limits of the whole enterprise.
There is no requirement for caffeine. Nobody develops a caffeine deficiency. It is not a nutrient.
The ergogenic evidence, by contrast, is about as good as nutrition evidence gets — measurable performance effects, replicated, with a dose-response relationship and objective outcomes rather than recalled ones.
The population data suggests benefit across a range of intakes. And then you meet actual people.
One cup and some people get palpitations. Others drink four in the evening and sleep fine. Some cannot open their bowels without a morning coffee; others become constipated from it. One cup triggers migraine in some people and nothing whatsoever in others.
All of that is real, and all of it is invisible in a study reporting an average.
The individual variation isn’t noise around the finding. Genetic differences in caffeine metabolism, adenosine receptor variants, habitual intake, iron status, medication, gut function and stress physiology all move the answer. A population average across thousands of people describes nobody in particular.
“How much coffee is good for you” has no answer at population level. It only has an answer for a person.
I read the study design before the conclusion. Was it a dish, an animal, or a person? Was diet measured or remembered? Was it randomised, or is this an association? Who funded it? Has anyone replicated it?
I take mechanism seriously as a reason to investigate and refuse to treat it as proof. Something can be biochemically elegant and clinically useless.
I pay attention to subgroups, because the omega-3 finding is not an anomaly. It’s how most of this works. Interventions help people who need them and do little for people who don’t, and averaging the two together produces a null result that means nothing about either group.
And I test rather than guess — not because testing is infallible, but because a measurement of the person in front of me beats an average derived from a questionnaire someone else filled in from memory.
Ask what kind of evidence it is before you ask what it says. In vitro, animal, observational, randomised — those are four different claims and only one of them establishes cause.
Treat mechanism as a hypothesis, not a finding. “It reduces inflammatory markers in cells” is a reason to look further. It is not a reason to buy anything.
Be suspicious of anything that works for everyone. Real interventions have responders and non-responders. A claim that admits no variation is usually a marketing claim.
And notice when the answer depends on who you are. Where it does — which is most of the time — the useful move isn’t reading more studies. It’s finding out where you actually stand.
None of this makes testing infallible. Laboratory reference ranges have their own methodological problems, and a marker is still a proxy for something you cannot observe directly. But a measurement taken from you is a different order of evidence from an average derived from other people’s recalled diets. The point of the TDG Five-Test Programme is not that it produces certainty — it is that it moves the question from what happens to people in general to what is happening in you.
If the answer depends on who you are, the useful next step is finding out. Ask AIdan which testing fits your situation, or start with a discovery call.
Ask AIdan → Book a Discovery Call